Agentic AI Runtime Stack

“SaaS is CRUD databases with business logic. As AI takes over that logic, SaaS will collapse.” -Satya Nadella

Today, Jake Flomenberg from Wing.VC and I published an Agentic AI Runtime Stack market map. We’re seeing the rapid proliferation of agents: software that uses large language models to reason through fuzzy situations to accomplish higher-level goals.

This market map was shaped by many conversations with agent builders and founders over the past few months, as well as my personal experiences in building an agentic system.

Some additional thoughts on the agents:

• A new term that I’m starting to use is agentic system: a network of agents. Because LLMs (and thus, agents) are stochastic in nature, I believe that decomposing high-level tasks & goals into smaller sub-agents is the right approach today. This makes it easier to ground and test each agent.

• A network of agents is … a distributed system. So all the strategies that were popularized by the cloud-native era—backpressure, retries, circuit breakers, distributed tracing, asynchronous processing—are even more important.

• I’m still pessimistic about agent frameworks, but I think LlamaIndex (with Workflow), LangChain (with LangGraph and LangSmith), and Mirascope (with Lilypad) are moving away from the “syntactic sugar” framework.

• Temporal and Akka aren’t standing still, either — they’re evolving to compete with the agent frameworks as well, just as the agent frameworks start to encroach on durable execution. Exciting times!

• Actuators. I don’t know if it will catch on, but the term “tools” is strangely generic and narrow at the same time. So we felt that tools was a poor choice to describe a big and exciting and rapidly changing category.

Thanks to Jake for the idea. It’s been great working with him. More to come!